Universal Joint Source-Channel Coding for Modulation-Agnostic Semantic Communication

I see uJSCC less as a new semantic objective and more as a deployment paper. Once semantic communication moves into digital modulation, the practical question is no longer only “Can a neural codec reconstruct the image?” It becomes “Do we need a different neural codec for every modulation order?” If the answer is yes, the system becomes hard to train, ship, update, and debug.

The paper’s answer is a shared encoder-decoder with a small modulation-specific branch in the statistics. The vector-quantization step still depends on the selected modulation path:

$$z_i^{(k)} = \arg\min_j \left\|y_i^{(k)} - c_j^{(k)}\right\|_2^2$$

The path itself is chosen from the SNR region:

$$m = m_k,\qquad b_{k-1} \le \eta < b_k$$

The part that makes the design economical is switchable batch normalization:

$$\mathrm{BN}_k(t) = \gamma_k\frac{t-\mu_k}{\sqrt{\sigma_k^2+\epsilon}} +\beta_k$$

Most of the model remains shared, while each modulation order keeps its own feature statistics. Training then ties the regimes together through a weighted loss:

$$\mathcal{L}_{\mathrm{total}} = \sum_{k=1}^{K}\lambda_k\mathcal{L}_k$$

This is a sensible compromise. Full sharing is too rigid because different modulation paths create different internal distributions. Fully separate models are expensive. S-BN sits between those extremes and makes model inventory smaller.

The weakness is that the SNR boundaries are still hand-designed. That is acceptable for a first system, but in a real link the best boundary depends on hardware, coding rate, mobility, channel estimation error, and the downstream semantic task. The method also assumes a known finite set of modulation modes. I would extend it by learning the boundary policy, then combining that policy with an information bottleneck objective. uJSCC can handle model sharing across modulation; RIB can decide how much redundancy should survive the channel.